@inproceedings{park-etal-2026-keep,
title = "What to Keep in Tight-Budget Dialogue Compression: Turn-Level Information Matters",
author = "Park, Minji and
Yoon, Seunghyun and
Lim, Hyuk",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.719/",
pages = "15123--15130",
ISBN = "979-8-89176-462-0",
abstract = "Dialogue histories often need to be compressed to fit within the context budgets of large language models (LLMs). However, when conversational compression occurs, turns containing critical information are sometimes removed. We study which turns should be retained to minimize information loss during task-oriented dialogue compression. On the MultiWOZ and SGD datasets, we compare four turn-level signals: Frame for task context, Evidence for question-conditioned relevance, Pivot for state changes, and Recency for later context. Evenly spaced selection (Stride) serves as a simple positional baseline.\par The effectiveness of each signal depends on the question type. Evidence achieves the highest accuracy among the four signals on current-slot questions in both datasets. At a 30{\%} word budget, Pivot improves over Stride on MultiWOZ when a state change determines the final value (0.421 versus 0.255), but is slightly worse on SGD (0.234 versus 0.245). For questions that require information from both the task context and a final value located in separate turns, Stride performs best among the compressed methods. The best selector depends on the evaluation target, and no single method is best in all settings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="park-etal-2026-keep">
<titleInfo>
<title>What to Keep in Tight-Budget Dialogue Compression: Turn-Level Information Matters</title>
</titleInfo>
<name type="personal">
<namePart type="given">Minji</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seunghyun</namePart>
<namePart type="family">Yoon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hyuk</namePart>
<namePart type="family">Lim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Dialogue histories often need to be compressed to fit within the context budgets of large language models (LLMs). However, when conversational compression occurs, turns containing critical information are sometimes removed. We study which turns should be retained to minimize information loss during task-oriented dialogue compression. On the MultiWOZ and SGD datasets, we compare four turn-level signals: Frame for task context, Evidence for question-conditioned relevance, Pivot for state changes, and Recency for later context. Evenly spaced selection (Stride) serves as a simple positional baseline.\par The effectiveness of each signal depends on the question type. Evidence achieves the highest accuracy among the four signals on current-slot questions in both datasets. At a 30% word budget, Pivot improves over Stride on MultiWOZ when a state change determines the final value (0.421 versus 0.255), but is slightly worse on SGD (0.234 versus 0.245). For questions that require information from both the task context and a final value located in separate turns, Stride performs best among the compressed methods. The best selector depends on the evaluation target, and no single method is best in all settings.</abstract>
<identifier type="citekey">park-etal-2026-keep</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.719/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>15123</start>
<end>15130</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T What to Keep in Tight-Budget Dialogue Compression: Turn-Level Information Matters
%A Park, Minji
%A Yoon, Seunghyun
%A Lim, Hyuk
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F park-etal-2026-keep
%X Dialogue histories often need to be compressed to fit within the context budgets of large language models (LLMs). However, when conversational compression occurs, turns containing critical information are sometimes removed. We study which turns should be retained to minimize information loss during task-oriented dialogue compression. On the MultiWOZ and SGD datasets, we compare four turn-level signals: Frame for task context, Evidence for question-conditioned relevance, Pivot for state changes, and Recency for later context. Evenly spaced selection (Stride) serves as a simple positional baseline.\par The effectiveness of each signal depends on the question type. Evidence achieves the highest accuracy among the four signals on current-slot questions in both datasets. At a 30% word budget, Pivot improves over Stride on MultiWOZ when a state change determines the final value (0.421 versus 0.255), but is slightly worse on SGD (0.234 versus 0.245). For questions that require information from both the task context and a final value located in separate turns, Stride performs best among the compressed methods. The best selector depends on the evaluation target, and no single method is best in all settings.
%U https://aclanthology.org/2026.findings-emnlp.719/
%P 15123-15130
Markdown (Informal)
[What to Keep in Tight-Budget Dialogue Compression: Turn-Level Information Matters](https://aclanthology.org/2026.findings-emnlp.719/) (Park et al., Findings 2026)
ACL